Model Training

Abstract

The chapter provides an approach and issues for model-training using Logistic Regression. (1) Regression—key model qualities plus i) options and settings, and ii) outputs to be expected/demanded. (2) Variable selection—i) criteria; ii) automation; iii) stepwise review; iv) constraining betas, where coefficients do not make sense; v) stepping by Gini, model pruning. (3) Correlation checks—i) multicollinearity—checks of variance inflation factors; ii) correlations—further checks to guard against the inclusion of highly correlated variables. (4) Blockwise variable selection—treatment in groups: i) variable reduction; ii) staged, or hierarchical regression; iii) embedded, model outputs as predictors; iv) ensemble, using outputs of other models. (5) Multi-model comparisons—Lorenz curves and strategy curves, should choices not be clear. (6) Calibration—i) simple adjustment by a constant; ii) piecewise, varying adjustments by the prediction; iii) score and points—adjusting the final score or constituent points; iv) MAPA, for more complex situations